A Study of the Air Pollution Index Reporting System

Transcription

1 Tender Ref. AP A Study of the Air Pollution Index Reporting System FINAL REPORT 27 June 2012 Submitted by Prof. Wong Tze Wai School of Public Health and Primary Care The Chinese University of Hong Kong (on behalf of the study team) Members of the Consultancy Team: The Chinese University of Hong Kong Prof. Wong Tze Wai (Principal Investigator and overall Project Manager) Dr. Wilson Tam Wai San (Model construction and statistical analyses) Prof. Yu Tak Sun, Ignatius (Epidemiological input and comments) Ms. Andromeda Wong Hin Shun (Editor) Hong Kong University of Science and Technology Prof. Alexis Lau Kai Hon (Literature review and comments) Mr. Simon KW Ng (Literature review) Mr. David Yeung (Computing) The University of Hong Kong Prof. Wong Chit Ming (Epidemiological input and comments) Citation of authorship: Wong TW, Tam WWS, Lau AKH, Ng SKW, Yu ITS, Wong AHS, Yeung D. 1

2 CONTENTS Page 1 Background 3 2 Objective 3 3 Literature Review Introduction AQI/API Construction Key Air Pollutants Averaging Times Calculation of AQI / API Reporting of AQI / API Comparing Different AQI / API Readings Recent Developments Air Quality Health Index in Canada Air Pollution Index System in South Africa Common Air Quality Index of European Union Comparison of API values using different levels of AQOs HK API based on WHO AQG HK API based on WHO AQG-NS HK API based on WHO AQG-F Comparing the Level of Exceedance Modelling hospital admissions data using the Canadian approach Rationale for the use of the Canadian model Statistical Modelling Banding of the Excess Risk of Hospital Admissions Attributable to Air 15 Pollution 6 Results Air pollutants and Emergency Hospital Admissions for Cardio-Respiratory diseases 6.2 Sensitivity Analysis Excess Risks of Hospital Admissions Attributable to Air Pollution Excess Risks of Hospital Admissions Attributable to Air Pollution Among High-Risk Groups 6.5 Health risk categories and AQHI Bands Interaction of air pollutants with cold season Annual Air Quality Index 23 7 Discussion Conclusion and Recommendation 27 9 Acknowledgement References Appendices Appendix 1: The problem of time lag for the API as an indicator of the current air pollution situation Appendix 2: Distribution of air pollutant concentrations by AQHI bands Appendix 3: Methodological issues in handling missing air pollutant data 34 Appendix 4: Plot of residuals against predicted hospital admissions in core model 35 Appendix 5: Plot of residuals against days 36 Appendix 6: Partial autocorrelation function by lag days 37 Appendix 7: Results from sensitivity analysis 38 Appendix 8: Viewpoints and discussions among team members on the Report Appendix 9: Comments by Health Canada on the Report Appendix 10: Response to Environment Canada s Comments

3 1 Background The Air Pollution Index (API) Reporting System is an important tool of risk communication. It informs the public of the local level of ambient air pollution, and the potential health risk it would impose, particularly on vulnerable groups such as children, the elderly, and those with existing cardiovascular and respiratory diseases. People use the API to help them make decisions on outdoor activities; for example, schools and sports organizations may check the latest API figures to decide whether outdoor sporting events should be conducted on a certain day. The Hong Kong API has been devised in a similar way to API systems used in other developed countries, although there are variations in the calculation methods. In June 2007, the Environmental Protection Department (EPD) of the Hong Kong SAR Government commissioned an 18-month study (Agreement No. CE 57/2006 (EP): Review of the Air Quality Objectives and Development of a Long Term Air Quality Strategy for Hong Kong Feasibility Study), to review Hong Kong s existing Air Quality Objectives (AQO), first established in 1987; and, following the review, to develop a long-term air quality strategy to achieve the updated objectives. This is in response to the Air Quality Guidelines (AQG), published by the World Health Organization (WHO) in October 2005 for worldwide adoption. As the calculation of the Hong Kong API is based on the 1987 AQO currently being reviewed, it is an opportune time to devise an improved API system that serves as an effective tool of risk communication to the general public. 2 Study Objective To develop an API reporting system for use in Hong Kong, with full justifications and implementation details. 3 Literature Review We have conducted a literature review of the API reporting systems in various countries, as stipulated in the Tender of this Study. 3.1 Introduction We reviewed the major air quality index (AQI) or air pollution index (API) systems around the world, including the United States (US), the United Kingdom (UK), Canada, Australia, China, France, Singapore, South Korea, Taiwan, South Africa, Macau, and Hong Kong (Ove Arup, 2007, 1 and 2; websites 1-14). While there are variations among the AQI / API systems developed by different countries or jurisdictions, all of the systems are designed to report the state of the air quality in a specific area or region, and to communicate its associated health risk. AQI / API systems are, in principle, designed to communicate the short-term health impact of local air quality to members of the public (Stieb et al, 2005), although in the US system, references are also made to long-term health risks. Health advisories are issued when the air pollution level is high, so that the general population, including susceptible groups, may take the necessary short- 3

4 term precautions. 3.2 AQI / API Construction In essence, an AQI or API is constructed to express the levels of one or more air pollutants, over various critical averaging periods, against a reference. The national air quality standards will usually be used as the reference for the index. A network of air quality monitoring stations will be set up to measure ambient concentrations of common pollutants at fixed intervals. Some monitoring stations are located at the roadside to measure street-level concentrations. In places like Hong Kong and Paris, a roadside or traffic index is reported separately from the general AQI / API (websites 3, 6) Key Air Pollutants There are variations with respect to the selection of key air pollutants, as individual countries or jurisdictions will seek to include pollutants that pose the most significant impact on their residents (Elshout & Léger, 2006). Air pollutants commonly used in AQI / API include nitrogen dioxide (NO 2 ), sulphur dioxide (SO 2 ), ozone (O 3 ), carbon monoxide (CO), respirable suspended particulate matter (PM 10 ), and lead. Fine suspended particulate matter (PM 2.5 ) is chosen in a few places, while in some Australian states, visibility is also incorporated into the AQI / API calculation (website 4). Conversely, pollutants that appear insignificant in a particular country may be omitted from the national AQI / API model. For example, Canada s Air Quality Health Index (AQHI) does not consider the concentrations of SO 2 and CO (Website 2). China s API excludes O 3 from its calculations (website 10) Averaging Times Another important aspect in constructing an AQI / API is the choice of the averaging time(s) for each pollutant. As the primary objective of an AQI / API system is to communicate the health risk related to short-term exposure to air pollutants, it would therefore be natural for the system to track pollutant concentrations over a shorter averaging time. Based on experience around the world, calculation of the AQI / API is usually based on 1-hour, 8-hour, or 24-hour average monitoring data, depending on the pollutants (Ove Arup, 2007, 1 & 2). It is worth noting that while the concentrations of most air pollutants are measured by a shorter averaging time (like the 1-hour average) for AQI / API calculations, particulate matter (PM) is averaged over a 24-hour period. This is due to the lack of scientific evidence with respect to the exposure-response relationship for PM over a one-hour period (Cairncross et al, 2007). As a result, when PM is the dominant pollutant, the AQI / API system is not responsive enough to reflect a sudden surge in the level of PM, because the index is based on its concentrations averaged over the past 24 hours. There is inevitably a time lag between the rise in concentration recorded at the monitoring stations and the rise in AQI / API readings; this time lag will delay the issuance of health advisories for impending air pollution episodes. An example that highlights this problem is presented in Appendix 1. 4

5 One possible approach to tackling this issue is to incorporate estimated pollutant concentrations for future hours into the calculation of the air quality index. The US has been predicting 8-hour ozone levels, based on the correlation between daily maximum 1-hour and 8-hour ozone values, in order to report AQI and health warnings in a more timely manner. Similarly, the AQI for PM is derived from the average of the past 12 hours and the predicted concentrations in the coming 12 hours (USEPA, 2006) Calculation of AQI / API Ambient or roadside concentrations for each pollutant, over different averaging times, will be converted into an index value. In general, there are three common methods to achieve this. The most popular approach is often called the US-based system. Pollutant concentrations for each pollutant are transformed onto a normalised numerical scale of 0 to 500, with an index value of 100 corresponding to the primary National Ambient Air Quality Standard (NAAQS) for each pollutant (USEPA, 2006, website 14). Places like Singapore, China, Thailand, Malaysia, South Korea, Taiwan, Hong Kong, and Macau designed their AQI / API systems based on the US model. The key reference point of these systems would be the index value of 100, which is based on the short-term air quality standards of the respective jurisdictions. Very often, the index value of 50 is anchored to the long-term air quality standards. A similar approach is being used in Australia, whereby pollutant concentrations are also being transformed onto a scale. There, however, a linear or proportional scale is used instead of a normalised scale (i.e. a scale which takes the variation into account), and the index is then calculated in direct proportion to the air quality standards or environmental goals (Ove Arup, 2007, 1). Moreover, the scale used in New South Wales is different from the one used in Queensland, Victoria, and Adelaide (in South Australia). In New South Wales, an index value of 50 means that the pollutant concentration is equal to the standard level. For the other states and cities, the index value of 100 carries the same meaning (Ove Arup, 2007, 1; website 4). The third approach is the banding system, which is more popular in European countries like the UK and France (websites 3, 13). The main deviation is that instead of using an index scale of 0 to 500, a scale of 0 to 10 is being used. For the UK system, this index scale of 10 is further broken down into four bands of low (1-3), moderate (4-6), high (7-9) and very high (10) (website 13). The key reference point for this banding system is the breakpoint value between the low and moderate bands. The lower bound of index value 4 is set to correspond to the UK Air Quality Standards for all pollutants but NO2. In this case, the 1-hour national standard for NO 2 is 200 g/m, whereas the lower bound of index value 4 for NO 2 is 287 g/m 3 (website 13) Reporting of AQI / API Based on one of the above three approaches, concentrations measured over various averaging times at individual monitoring stations will be transformed into air pollution sub-indices (APSI) for each of the pollutants. Normally, the 5

6 highest of the sub-indices will be taken as the reported AQI / API, and the contributing pollutant will also be specified. Reporting the air quality as designated by the level of the single worst pollutant has its limitations. In the real world, multiple pollutants affect the health of the community simultaneously, and the conventional approach simply ignores the joint effects of different air pollutants on human health. For instance, we would logically expect a greater impact on health when several pollutants are breaching their respective short-term standards at the same time, as compared to one pollutant reaching an unhealthy concentration level on its own (Cairncross et al, 2007). However, the simple addition of the health risks of each air pollutant derived from single-pollutant models, as in the case of Canada s AQHI (see section below), may be an over-representation of the total health effect, by assuming the effects of each pollutant are independent of the others and the total effects are the sum of the individual effects. While some studies have shown that certain pollutants might have synergistic effects, it is not impossible that some pollutants might antagonise the effect of another. How to assess the joint health risks of multiple air pollutants will remain a subject of debate and future research. In some places, such as China, a different approach is taken whereby the daily average of a pollutant concentration at a monitoring station will be derived from the hourly readings, and a sub-index will then be calculated for that pollutant. The highest sub-index of the most critical pollutant will become the AQI / API of the area (website 10). For effective communication, descriptors, colour codes, and health advice or warnings are often assigned to specific ranges of AQI / API values. However, there is no universal guideline regarding the wording of the descriptors or health advisories, or on the colour scheme to be used Comparing Different AQI / API Readings Comparing the air quality in different countries using AQI / API readings is always a difficult endeavour. Firstly, arguably few AQI / API systems are identical. Individual country and jurisdictions will design their own systems to report local air quality in the most appropriate way, which means they would choose different air pollutants (those that predominantly affect the local population) and different reporting systems (using an index scale or a banding system). Secondly, air pollutant concentrations are often measured at different locations within a city that are not directly comparable. For instance, air quality indices representing measurements taken from the ambient air at background stations are very different from those taken from roadside stations, which are influenced by traffic (Elshout & Léger, 2006). Thirdly, even for the same measured pollutant concentration, different countries may have different interpretations with respect to its health effect and additional health risk (Elshout & Léger, 2006). For example, in France, the worst endpoint ( very poor ) of the NO 2 sub-index is 400 g/m 3 (website 3). In the UK, the same value is taken as the lower end of the moderate band (website 13). 6

7 The AQI / API systems are, in many ways, a gross generalization of a complex mixture of airborne chemicals into a simple index value. The primary purpose for which they are designed is risk communication to the public, rather than comparison between different cities. 3.3 Recent Developments Air Quality Health Index in Canada Canada has been using an Air Quality Index (AQI) system to report current and near-term air quality conditions. A scale of 0 to 100 represents air quality conditions ranging from very good to very poor (website 2). An air quality advisory is issued when the calculated sub-indices of the pollutant concentration exceed, for a fixed period of time, an AQI value of 50, at which point the air quality is defined as changing from moderate to poor (website 2). While the AQI remains a simple tool for communicating the state of the local air, there is little national consistency in how AQIs are reported. The pollution thresholds, the pollutants included in the AQI formulation, and the use of healthbased messages vary from one place to another across the country (website 2). Notably, the thresholds used in determining AQI levels and categories are often based on outdated health science, and tend to reflect environmental regulatory imperatives rather than implications for human health (website 2). In June 2001, the Government of Canada began working with a variety of stakeholders to address the shortcomings of their conventional AQI system, and to devise an effective risk communication tool for acute health effects. Inadequacies of the conventional system included (a) its failure to consider the combined effects of multiple pollutants; (b) its failure to reflect the no-threshold concentration-response relationship between air pollution and health; and (c) its linkage with standards that might be influenced by factors other than health risk (Stieb et al, 2008; Taylor, 2008; website 2). A new Air Quality Health Index (AQHI) has been designed to help people understand what a certain state of local air quality means to public health. A national pilot programme began in July 2007 for the city of Toronto. At present, the AQHI is available for about ten communities in Canada, including Vancouver and Victoria (website 2). The AQHI is constructed as the sum of excess mortality risk associated with NO 2, ground-level O 3, and PM 2.5 at certain concentrations. It is calculated hourly based on 3-hour rolling average pollutant concentrations, and is then adjusted to a scale of 1 to 10. The value of 10 corresponds to the highest observed weighted average in an initial data set, measured in 10 Canadian cities and covering the period between 1998 and 2000 (Stieb et al, 2008; Taylor, 2008). The scientific foundation for the AQHI is based on the epidemiological research undertaken at Health Canada. Relative risk (RR) values are estimated, based on local time-series analyses of air pollution and mortality (Stieb et al, 2008; Taylor, 2008). The AQHI index values are grouped into four health risk categories: low (1-3), moderate (4-6), high (7-9) and very high (10+). Health messages customized to each category, for both the general population and the at risk population, will be disseminated (Stieb et al, 2008; Taylor, 2008; website 2). 7

8 3.3.2 Air Pollution Index System in South Africa A similar health-based index has been developed in South Africa in a dynamic air pollution prediction system (DAPPS) project, which is led by a consortium of four South African partners, including the Cape Peninsula University of Technology (Cairncross et al, 2007). This API system is based on the relative risk of the well-established excess daily mortality associated with short-term exposure to common air pollutants, including PM 10, PM 2.5, SO 2, O 3, NO 2 and CO. A set of relative risks published by the World Health Organization has been used to calculate sub-index values for particulates, SO 2, O 3 and NO 2. For CO, an RR value of 1.04 (for a 10 ppm increment in exposure) was used after Schwartz (1995). O 3 concentrations in the WHO guidelines was used as a reference level for mortality risk, which forms the basis for calculating the concentrations of other pollutants, A scale of 0 to 10 is used. Incremental risk values for each pollutant are assumed to be constant, and a continuous linear index scale is developed for each pollutant, with RR = 1 at zero exposure. For consistency between pollutant exposure metrics, the exposures that correspond to the same relative risk are assigned the same sub-index value. The final API is the sum of the normalised values of the individual indices for all the pollutants. The proposed API has been applied to ambient concentration data collected at monitoring stations in the City of Cape Town for testing. However, it is unsure whether the system has been put into any pilot programme in South Africa. Following the method by Cairncross et al (2007), Sicard et al. (2011), developed an aggregate index using five air pollutants (PM 2.5, PM 10, NO 2, O 3 and SO 2 ) for the Provence Alpes Côte d Azur (PACA) region, in the South East of France, using PM 2.5 as a reference instead. This aggregate index will be used in three European sites Greece (Athens and Thessaloniki), the Netherlands, and PACA region (Sicard et al, 2012) Common Air Quality Index in the European Union The Common Air Quality Index (CAQI) has recently been developed by the European Union. Three different indices hourly, daily and annual present the air quality conditions in European cities in a simple and comparable way. Both background and roadside situations are represented. The hourly and daily indices are expressed using a 5-level scale, ranging from 0 (very low) to >100 (very high). The calculation is based on concentrations of PM 10, NO 2, and O 3, which are the three pollutants that raise major concerns in Europe. The indices reflect EU alert threshold levels or daily limit values as much as possible. The annual index, on the other hand, provides an overview of the air quality situation in a given city throughout the year, with respect to the EU standards. It is developed to reflect the effect of long-term exposure to air pollution. The annual index is presented as a comparison to the EU annual air quality standards and objectives. If the index value is higher than 1, the limit values of one or more pollutants are not met. If the index value is below 1, on average all the limit values are met. 8

9 It is important to note that the CAQI is a standards-based system and is designed to give a dynamic picture of the air quality situation in each city but not for compliance checking (website 7). 4 Comparison of API values using different AQO A password-protected webpage ( has been created to calculate the would-be index values by adopting the reporting systems from different countries / states / cities, and to compare them with Hong Kong s current API. The systems included here are China s API, Macao s API, Taiwan s Pollution Standard Index (PSI), USA s AQI, South Korea s Comprehensive Air-quality Index (CAI), Ontario s AQI, British Columbia s AQI, New South Wales AQI, UK s API, Canada s AQHI, and a derivation of Canada s AQHI with FSP approximated as 0.7 RSP. In addition, three more sets of indices, each using the Hong Kong API calculation methodology but with different subindex thresholds one based on the WHO AQG and the other two, some modifications of the WHO AQG have also been calculated. 4.1 HK API based on WHO AQG This is the most direct application of the WHO AQG (numbers shown in red and italics), with linear interpolation below the guideline and linear extrapolation above the guideline (Table 1). For monitoring stations without PM 2.5 measurements, we estimated the values from PM 10 by the formula: PM 2.5 = 0.7 PM 10. Table 1: Hong Kong APSI as calculated by WHO AQG APSI* PM hr SO 2 24-hr NO 2 1-hr NO 2 24-hr O 3 8-hr PM hr *APSI: Air pollution sub-index 4.2 HK API based on WHO AQG-NS As we apply the WHO AQG directly to the APSI thresholds as defined above, SO 2 24-hr becomes the dominant contributing pollutant in general stations most of the time (Table 6). This differs from the findings of most health studies, which suggest that other pollutants (e.g. PM and NO 2 ) are more important than SO 2 in terms of overall health risk to the public. Hence, another set of APSI thresholds, known in our plots as WHO AQG-NS, was defined by taking away the SO 2 thresholds from the calculation. The thresholds are shown in Table 2. The PM 2.5 levels for stations without PM 2.5 measurements were estimated using the same formula as before. 9

10 Table 2: Hong Kong APSI as calculated by WHO AQG without SO 2 APSI PM hr NO 2 1-hr NO 2 24-hr O 3 8-hr PM hr HK API based on WHO AQG-F The final index was defined by adding arbitrary hourly thresholds for PM 10, PM 2.5 and O 3 to help minimize the timedelay / phrase shift problem of the currently defined API. The index was identified as WHO AQG-F, and the thresholds are shown in Table 3. In this model, the short-term (1-hour) thresholds were arbitrarily set to be double that of the longer-term (24-hour or 8-hour) thresholds. Again, the PM 2.5 levels for stations without FSP measurements were estimated by the same formula as before. Table 3: Hong Kong APSI as calculated by WHO AQG with hourly PM 10, PM 2.5 and O 3 APSI PM hr PM 10 1-hr NO 2 1-hr NO 2 24-hr O 3 8-hr O 3 1-hr PM hr PM hr Comparing the Level of Exceedance When we calculated Hong Kong s API using methods from different reporting systems and different thresholds, the number of days of exceedance (on which API > 100), as well as the relative significance of each air pollutant as the contributor to the daily maximum, varied from one method to another. In Table 4, we have listed out the number of days of exceedance based on Hong Kong s current API calculations by contributing pollutants. It is clear that NO 2 has been the major contributor towards non-compliance at roadside stations, while O 3 has been responsible for most of the exceedance days at the general stations. SO 2 never contributed to any instances of exceedance during this period, while PM 2.5 was not a contributor because we have no AQO for this air pollutant. 10

11 Table 4: HK API based on AQO: Number of Days of Exceedance HK AQO Number of Days Exceeding HK AQO as the Contributing Pollutant PM 2.5 NO 2 O 3 PM 10 SO 2 Total 2000 General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside We then compared the level of exceedance based on the API (daily maximum) calculated from four different sets of threshold values, namely the WHO AQG interim target 1 (IT-1) thresholds, the WHO AQG ultimate values, the WHO AQG-NS values, and the WHO AQG-F values. Table 5 shows that with WHO AQG IT-1, the number of exceedance days increased significantly. Amongst the pollutants, the contributions of NO 2 and PM 2.5 to non-compliance have also increased significantly, both at the general and roadside stations. On the other hand, the number of exceedance days due to O 3 dropped slightly. SO 2 contributed very infrequently to exceedance, while PM 10 never contributed at all. Table 5: HK API based on WHO AQG IT-1: Number of Days of Exceedance WHO AQG IT-1 Number of Days Exceeding AQG IT-1 as the Contributing Pollutant PM 2.5 NO 2 O 3 PM 10 SO 2 Total 2000 General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside

12 When we set the WHO AQG s ultimate limits as the threshold values for calculation (Table 6), there were significant changes in the API. Firstly, exceedance occurred almost every day. This was expected, as the threshold values for all pollutants are very stringent under the WHO AQG. Secondly, PM 2.5 became even more dominant as the contributing pollutant at the roadside stations. Thirdly, SO 2 emerged as a significant contributor to non-compliance with the 24-hour standard, especially at the general stations. Fourthly, the role of NO 2 became overshadowed by the other pollutants. However, it is important to note that the concentrations of NO 2 were still fairly high, even though the gas was relatively less dominating than pollutants such as FSP or SO 2. Table 6: HK API based on WHO AQG: Number of Days of Exceedance WHO AQG Number of Days Exceeding AQG as the Contributing Pollutant FSP NO 2 O 3 PM 10 SO 2 Total 2000 General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside Tables 7 and 8 show the number of exceedance days by applying the threshold values of WHO AQG-NS (see Table 2) and WHO AQG-F (Table 3). The results were quite similar in both cases, with PM 2.5 increasing in dominance at the expense of SO 2. Table 7: HK API based on WHO AQG-NS: Number of Days of Exceedance WHO AQG-NS Number of Days Exceeding AQG-NS as the Contributing Pollutant PM 2.5 NO 2 O 3 PM 10 SO 2 Total 2000 General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside

13 Table 8: HK API based on WHO AQG-F: Number of Days of Exceedance WHO AQG-F Number of Days Exceeding AQG-F as the Contributing Pollutant PM 2.5 NO 2 O 3 PM 10 SO 2 Total 2000 General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside General Roadside Modelling Hospital Admissions Data using the Canadian Approach 5.1 Rationale for the Use of the Canadian Model In the literature review, we found that the AQHI adopted by Canada (Stieb et al, 2008) was a unique, health-based system that made use of local health data, established a link with air pollution data, and then calculated the impact of different levels of air pollution on a specific health outcome (mortality). The purpose was to ensure that the API reporting system will be based on health outcomes observed locally, instead of on study findings in other countries, where the quantitative relation between air pollution and health might be different and therefore not directly applicable. Another feature of the Canadian AQHI system was that it combined the effects of multiple pollutants, assuming them to be independent and hence additive. The use of a daily maximum of the 3-hour moving average in the construction of the statistical model was a compromise between timeliness (using real-time data) and the delayed, cumulative effects of continuous exposure to air pollution. 5.2 Statistical Modelling We modified the Canadian system by substituting mortality data which are less sensitive indicators of health with emergency hospital admissions for respiratory and cardiovascular diseases. The advantage of using this indicator of illhealth is that we have a comprehensive and uniform dataset in the public hospitals. The data are subject to stringent quality control, and represent over 90% of all emergency hospital admissions throughout Hong Kong. We used local data to obtain relative risks (RR) for individual air pollutants. Assuming a linear dose-response relationship, and a zero excess risk when the air pollutant concentration reaches zero (rather than using an arbitrary standard like the WHO AQG as the zero excess risk reference), we calculated the proportion of excess emergency hospital admissions for respiratory and cardiovascular diseases that were attributable to air pollution at different levels of air pollution. 13

14 To estimate the RR, which quantifies the risk of hospital admissions for different air pollutants, either singly or in their joint effects, we performed a time series study using Poisson regression. Data on hospital admissions for respiratory and cardiovascular diseases (from 2001 to 2005) were obtained from the Hospital Authority. Daily meteorological variables (mean temperature and humidity) were obtained from the Hong Kong Observatory. The statistical model chosen was a generalized additive model, one that has been most widely used in the current literature. Daily emergency hospital admissions for respiratory and cardiovascular diseases were used as the health outcome variables in the model, and smoothing for the time variable was done for various degrees of freedom using smoothing splines. The model was adjusted for daily mean temperature and relative humidity, a day of the week indicator, a holiday indicator, and a season indicator as potential confounders. Over-dispersion was adjusted by the quasi likelihood method and autocorrelation was adjusted by adding auto-regressive terms into the core model. Residuals plots and PACF plots were used to examine the goodness of fit of the model. Hourly air pollutant concentrations were provided by the Environmental Protection Department, and the maximum of the 3-hourly moving average of a day was used to define the daily concentrations for each air pollutant. The model was tested for the lag effect of air pollution, on the same day (lag day 0), lag day 1 (air pollutant concentration on the previous day) and lag day 2 (two days ago). The best lag day for each air pollutant was chosen according to the maximum t-value, calculated using the gam.exact function of S-PLUS (ihapss, 2002; Dominici et al, 2004). There is no consensus in the literature on the choice of models for hospital admissions. For mortality data, investigators of the National Morbidity, Mortality, and Air Pollution Study (NMMAPS) used 7 degrees of freedom per year (Peng et al, 2005). The percentage of excess daily hospital admissions, also known as the percentage excess risk (%ER), was expressed as: βixij i=1,...,p (e -1) 100%. β i was the regression coefficient of pollutant i from the time series analysis, and x ij was the concentration of pollutant i at time j, for a total of p pollutants. The %ER was calculated using the regression coefficients of four pollutants, NO 2, O 3, PM 10 and SO 2 **, and their respective concentrations on each day over the five-year period. An equal weight was assumed for each pollutant. In the generalized additive model, the degrees of freedom (df) for the smoothing parameter tested were: 0, 10, 20,.to 160. RR estimates varied with varying df, and the RR values peaked at different df. NO 2 peaked at df=50; PM 2.5, at df=60, and O 3 at df=80, while SO 2 peaked at df=0. The model with 70 degrees of freedom was chosen in favour of the model with the minimum AIC (df=147) because the RRs of most air pollutants (except SO 2 ) were near their maximum values. For SO 2, the RR was much smaller than all the other pollutants and contributed little in the calculation of the %ER. At df=70, the relative weights of the RRs of all pollutants in their contribution to the %ER are much more balanced than that using RRs derived in the statistically best-fit model, where the relative weightage of O 3 is much higher than all other pollutants. The relative weightage of the beta values for NO 2, O 3, SO 2 and PM 10 at df=70 were: 32.4%:37.1%:10.1%:20.5%. At df=147, the relative weightage for NO 2, O 3, SO 2 and PM 10 were: 25.9%:51.4%:6.0%:16.7%. This implies the %ER will be dominated by O 3 concentrations while other pollutants contributions to the %ER are insignificant. The rationale for our model choice is to estimate the highest levels of risk that can be attributed to most air pollutants, while maintaining a more balanced contribution of %ERs among the 4 pollutants. It is generally recognized that PM 2.5 penetrate the lung more deeply than PM 10 do. However, in the calculation of excess hospital admissions, we have chosen PM 10 instead of PM 2.5, because data on PM 10 were available in all air monitoring stations, compared to PM 2.5, with data limited to a few stations only. The RR for PM 10 is therefore more robust than that for PM 2.5. It is anticipated that more comprehensive monitoring of PM 2.5 will be implemented by EPD in the future. In Hong Kong, PM 2.5 is strongly correlated with PM 10, with a high PM 2.5 to PM 10 ratio of about 0.7. ** The Canadian AQHI excluded SO 2 from its model. 14

15 Similar models were constructed for hospital admissions for children below 5 years of age and for those aged 65 years and above, and the corresponding %ERs were estimated using the respective β for the pollutants obtained from these models, as described above. To test the validity of the model, we performed sensitivity analyses by adding an indicator for influenza (Wong et al, 2002), and by splitting the 5 years time series into two ( and ) to estimate the RRs of the respective air pollutants. 5.3 Banding of the Excess Risk of Hospital Admissions Attributable to Air Pollution The excess risks were categorized into five bands, in terms of the risk level from short-term exposure to air pollution: Band 1 (low risk), Band 2 (moderate risk), Band 3 (high risk), Band 4 (very high risk) and Band 5 (serious risk). This banding used the short-term exposure limit values of the four air pollutants, as recommended by the World Health Organization s Air Quality Guidelines 2005 (with some modifications for NO 2), as reference points to identify a very high risk band one where the general public is exposed to a significant health risk. The rationale is that on a day with concentrations of air pollutants at the respective levels, the sum of %ER will be considered as the threshold above which the risk is too high. To address the health risk to the vulnerable groups children aged under 5, and the elderly aged 65 years and above, the reference point for the %ER of these groups was further adjusted. An adjustment factor was derived from the ratio of the median %ER for children under 5 or those aged 65 years and above (whichever was the larger) to that for all ages. The adjusted %ER, used as a limit for short-term exposure to air pollutants by the high risk groups, was obtained by dividing the %ER for all ages by the adjustment factor. Half of this adjusted %ER value was arbitrarily used as a dividing line between the low risk category (<0.5 adjusted %ER) and the moderate risk category (>0.5 adjusted %ER to < adjusted %ER). The %ER above the adjusted %ER up to the unadjusted %ER was categorized as high risk, as the air pollutant concentration would pose a significant health risk to the high-risk age groups but not to the other age groups. When the %ER is above the unadjusted %ER value, the health risk was categorized as very high, because the air pollutants would pose a significant health risk to people of all ages. A %ER 50% higher than the unadjusted %ER value is labelled as serious. The WHO Guidelines set the 24-hr mean for PM 2.5 and PM 10 as the values representing the 99 th percentile of the distribution of daily values, based on the relation between the daily mean and the respective annual mean AQGs. The 8-hr mean for O 3 and one-hour mean for NO 2 were derived from studies of short-term health effects, including time series and toxicological studies. The 24-hr AQG for SO 2 was based on an intervention study and a Hong Kong-London comparative study, but not on the relation between the distribution of the daily and annual mean concentrations, with an unusual, identical AQG for 24-hr and annual SO 2. The WHO AQGs for short-term exposure are: 200 g for NO 2 (one-hour), 100 g for O 3 (8-hour mean), 50 g for PM 10 (24- hour) and 20 g for SO 2. Since the averaging time for NO 2 was one hour, we calculated the corresponding value of the NO 2 concentration for a 3-hour moving average, by regressing the one hourly concentrations with the 3-hour moving average in a linear regression model using NO 2 data in our study period. The corresponding value was g/m 3 with a 95% lower confidence limit of g/m 3. The lower 95% confidence limit of g/m 3 was used as the concentration of our calculation of %ER for NO 2. 15

16 6 Results 6.1 Air Pollutants and Emergency Hospital Admissions for Cardio-Respiratory Diseases All five air pollutants were significantly associated with emergency hospital admissions for respiratory and cardiovascular diseases for all age groups combined: NO 2, O 3, PM 10, PM 2.5 and SO 2. Their respective RRs were: , , , and (Table 9). The RRs were significant at p< for the first four pollutants, and p= for SO 2. The best lag day was lag day 0 (same day) for all pollutants except O 3 (where lag day 1 was the best lag day ). The core model (before the air pollutant concentration was added) is shown as follows: Log (resp card 0) = resp card s (day, 70) + s (humidity, d.f. = 15) + s (temperature, d.f. = 15) + day of week indicator + season indicator + holiday indicator The residual plots did not show any obvious cyclical patterns (See Appendices 4 and 5). The PACF plots showed that autocorrelation was insignificant up to lag day 12 (Figure 2 of Appendix 6). Explanatory note: The dependent (outcome) variable, resp card 0, is the daily number of emergency hospital admissions for respiratory diseases and cardiovascular diseases. The independent variables are: Resp card 1 + resp card 2 + resp card 3 + resp card 4 + resp card 5 + resp card 6 + resp card 7 are the numbers of emergency hospital admissions for respiratory diseases and cardiovascular diseases from lag day 1, day 2, to day 7. They are also called auto-regressive terms. s (day, 70) is the time (or day) variable and is smoothed with 70 degrees of freedom (d.f.). s (humidity, d.f. = 15) is the daily mean humidity and is smoothed with 15 d.f. s (temperature, d.f. = 15) is the daily mean temperature (in Celsius) and is smoothed with 15 d.f. Day of week indicator shows the day of the week variable (Monday, Tuesday,, Sunday). (Cold) season indicator takes the value 1 from December to February and 0 during the period from March to November. Holiday indicator takes the value of 1 on public holidays. The RRs for high risk groups, namely those aged 65 years and above, and children under 5 years, were estimated using the same model. Compared to RRs for all ages, those aged 65 years and above had higher RRs for NO 2, PM 10, O 3 and SO 2, but slightly lower RR for PM 2.5. All RRs were significantly higher than one. The RRs for children under 5 years were even higher for O 3, PM 2.5 and SO 2, but lower than the RR for all ages for PM 10 and NO 2. The RRs were significant for O 3 and NO 2, but insignificant for PM 2.5 and SO 2. 16

17 Table 9: Relative risk of hospital admissions for cardiovascular and respiratory diseases per 10 µg/m 3 increase in air pollutant concentrations RR (95% CI) per 10 µg/m 3 increase in air pollutant concentration (single pollutant model) Emergency hospital NO 2 O 3 PM 10 PM 2.5 SO 2 admissions Cardiovascular and * respiratory (all ages) ( ) ( ) ( ) ( ) ( ) (d.f.=70) (lag day 0) (lag day 1) (lag day 0) (lag day 0) (lag day 0) #: Cardiovascular ** * and respiratory ( ) ( ) ( ) ( ) ( ) (>65 years) (lag day 0) (lag day 1) (lag day 0) (lag day 2) (lag day 0) (d.f.=70) Cardiovascular and ** * (NS) (NS) respiratory: ( ) ( ) ( ) ( ) ( ) (<5 years) (lag day 2) (lag day 0) (lag day 2) (lag day 1) (lag day 1) (d.f.=70) # The >65 years age group constituted about 80% of all respiratory and cardiovascular admissions. * p<0.05; ** p<0.001; p<0.0001; NS = not significant at p=0.05; d.f. = degree of freedom for the variable days 6.2 Sensitivity analysis To examine the effect of influenza on hospital admissions, we added an indicator variable on to the model using an arbitrary definition of an influenza week, as one during which the number of influenza hospital admissions exceeded the 75th percentile for the year (Wong et al, 2002). The differences in RR from that in our original model ranged from % to %, which had little effect on our calculation of % excess risks. There was little change in the statistical significance of the RRs (see Appendix 7). To test the stability of the model, we split the time series into 2 periods: and and ran separate models. All the RRs were similar to that in the original model, with differences ranging from % to 0.144%, for the 2-year model, and from % to % for the 3-year model. The 95% confidence intervals of the RRs in the split models were wider, but remained statistically significant, except for SO 2 in the 3-year model (see Appendix 7). 6.3 Excess Risks of Hospital Admissions Attributable to Air Pollution The frequency distribution of the daily excess risk of hospital admissions attributable to air pollution (expressed as a percentage, %ER) during the time period is shown in Figure 1. During the five years study period, the minimum of the percentage of excess daily hospital admissions attributable to air pollution (% ER) was 2.64%; the maximum was 31.51%, with a median of 9.04% and a mean of 9.50%. 6.4 Excess Risks of Hospital Admissions Attributable to Air Pollution Among High-Risk Groups The % ER for hospital admissions for cardiovascular and respiratory diseases were calculated for those aged 65 and above (with a minimum % ER of 3.02%, a median of 10.34%, a maximum of 36.25% and a mean of 10.86%) and children under 5 years of age (with a minimum % ER of 2.59%, a median of 9.44%, a maximum of 33.32% and a mean 17

18 of 10.01%). 6.5 Health risk categories and AQHI Bands The anchor point of the %ER that separated very high from the high band was derived from the sum of %ER values calculated from the recommended short-term exposure limit values of the WHO AQG for four pollutants: NO (modified), O 3, PM 10 * and SO2. These concentrations were: g/m for NO 2 (see footnote of section 5.3), 100 g/m for O 3 (8-hour mean), 50 g/m 3 for PM 10 (24-hour mean), and 20 g/m 3 for SO 2. Using these values, a %ER of 12.91% was obtained for all age groups. Above this %ER, the AQHI was considered unsafe even for healthy persons in the community and was labelled very high. A %ER above 19.37% (50% higher than 12.91%) was labelled as serious. For the high risk age groups comprising both children aged under 5, and the elderly aged 65 years and above the %ER was adjusted downwards by a factor of The %ER of 12.91% was divided by 1.144, giving a value of 11.29%. The %ER of 11.29% was used as an upper limit for short-term exposure to air pollutants by the high risk age groups. Above 11.29%, up to 12.91%, the air pollutant concentration was considered to pose a significant health risk to the high risk age groups, but not to the general population. This range of %ER was categorized as high. Half of 11.29%, i.e. 5.64%, was used as the cut-off point of low risk / moderate risk category. The %ER in this range (>5.64%, up to 11.29%) was labelled as moderate, while the %ER at 5.64% or below was labeled as low. The %ER in the categories low, moderate and very high were further sub-divided into equal thirds, making a total of 10 bands from low 1 to very high 10. The category serious is labelled as band 10+. The distribution of the %ER in the 5-year study period is shown in Table 10 and in Figure 1. The health advice corresponding to each band is shown in Table 11. * PM 10 was used in the calculation of %ER instead of PM 2.5 because data on the former are more comprehensive. In addition, the concentrations of PM 10 increase to a much greater extent than that of PM 2.5 during dust storm episodes, making PM 10 a better indicator of health impact on these days. The median %ER for the elderly was 10.34%, which was higher than that for the children (at 9.44%). The %ER for all ages was 9.04%. The ratio of the median %ER for the elderly to that for all ages was 10.34/9.04, or The ratio of the median %ER for the children to that for all ages was Hence, the adjusted %ER was 11.29% was used, as it was lower than that for children (at 12.37%). In response to demand from different sectors of the community, the health advice included persons with cardiovascular and / or respiratory diseases and outdoor workers, in addition to children and the elderly, and the general public. It should be noted that the advice was not based on the risk estimates from the results of this study. Advice for persons with cardiovascular / respiratory diseases generally follows that for children and the elderly, whereas outdoor workers were assumed to be healthy, non-elderly adults. 18